Robotization of Synthesis and Analysis Process of Graphene Oxide‐Based Membrane

The use of collaborative robots (Cobots) for materials development in chemical laboratories is currently of high priority. Herein, the Cobot is used for autonomous continued analysis and synthesis of graphene oxide–polyethyleneimine‐based membrane to unify a method and prospects for big data collection are shown. Membranes have already demonstrated a selective affinity to potassium cations and promised to adjust permeability for other cations by changing pH. The Cobot allows a variation of membrane properties by its composition modification. The present strategy combines a novel perspective of material production by Cobots and the application of machine learning. Moreover, the current approach can be adapted for different modern chemical laboratories for various scientific research, and the proper workflow is provided.


Introduction
The robotization of chemical laboratory processes is currently straightforward, for example, a multichannel automatic pipette development. [1]Together with other approaches, it allows scaling the experiments.Using robotic manipulators in humaninhabited environments such as chemical laboratories became possible with the rapid development of collaborative robots (Cobots).Unlike industrial robotic arms, Cobots are safe and friendly for humans, where external safety measures are needed: fences, light curtains, and mandatory emergency stop buttons. [2]Humans could not work in a shared environment with industrial robotic arms, but they could work with Cobots simultaneously.Cobots are designed to coexist with humans.Particularly, it has features such as builtin torque sensors and power limitations. [3,4]n chemical laboratories, scientists usually must operate expensive instruments, so using Cobots would also reduce the risk of damaging fragile equipment.Last year, A-lab was presented by Ceder Group. [5]hey use theory-driven and data-driven machine learning (ML) techniques to discover new materials with minimal human input.The systematic approach of the A-Lab allows for the generation and collection of data that can inform the system and the wider community.
In most recent research, the ABB YuMi collaborative robot is used to automate solid dispensing tasks. [6]Authors describe how they transferred actions previously performed by a researcher to be made by a Cobot, which, as they claim, would allow scientists to focus more on mentally stimulating tasks and reduce the risk of exposure to dangerous chemicals.In the article related to the recent COVID-19 outbreak, [7] authors propose using a Cobot in washing and pipetting actions during serological tests in understaffed facilities.The results show that utilizing Cobots in partial automation of chemistry tasks provides implementation and automation flexibility compared to the total laboratory automation.The autonomous Kuka platform does photocatalysis optimization very efficiently and improves the photocatalyst. [8]More than 500 experiments were performed during the 8 days.Autonomous research helps to find photocatalyst mixtures six times more effective than initial compounds.There is the possibility of using several collaborative, mobile robotic platforms to perform various chemical tasks. [9]Simultaneously, platform optimization is needed for each individual task.In the present work, we suggest such optimization for membrane materials.Open code is also provided.
[11][12] Last year, an open-source Deep Docking protocol was published as a new strategy for searching for a new drug without losing any potential possibility and to save the quantity of the material screening. [10]Many automatic features are used in molecular biology [13,14] or bacteriological tests [15,16] when samples should reproduce under the same conditions.Several research groups worldwide are applying mechanical autonomous systems.Cronin's group suggested a novel way to adjust the experimental parameters of the system and find the best catalyst among others. [17,18]Researchers defined a chemical reaction as several modules and developed "Reactionware" software.The software allows the standardization of the processes and increases their efficiency.Also, 3D-printed chemical vessels make the application area worldwide.
ML usage allows predicting properties of a lot of new materials, for example, membranes.Membranes play a fundamental role in many applications, such as filtration and recovery of resources. [19,20]That is why past decades were dedicated to "smart" membranes that can change properties with external stimuli by demand. [21,22]The development of intellectual materials increased membrane application and made a flexible permeability and separation performance. [23,24]In recent research graphene oxide-polyethyleneimine (GO-PEI) membrane with cation-selective properties was developed. [25]It was shown that pH strongly influences membrane permeability, especially potassium cations transfer through materials.It increases 20 times in acid media.The synergetic effect of pair cations (K þ and Cs þ ) was obtained: Cs þ began to pass through only by the presence of K þ .Interaction between ions inside the membrane may potentially lie in the permeable tuning.We understand that ML helps predict a property before a time-consuming experiment.
ML was carried out in the present work based on an experiment in our laboratory.Here, we use Cobot and ML to obtain selective GO-based membrane properties.The Cobot operates autonomously over 24 h, performing experiments.A fixed robotization station produces a GO-PEI membrane using the recently published methodology. [25]This modular approach could be applied to creating materials with different functionalities, such as selective permeability or accumulation of cations.
In this work, we designed a system to automate all synthesis stages, apply ML to resulting data, and make decisions concerning membrane properties.We focused on developing a platform that performs experimental steps, has a range of input reagents, and carries out recursive reaction cycles.

Results and Discussion
Our setup is shown in Figure 1a,b.The operation can be seen if one follows the QR code in Figure 1c.The experimental setup includes five principal blocks: 1) mixing of reactive; 2) centrifugation; 3) vacuum filtration; 4) drying, and 5) analysis.3D models of additional station equipment are fully represented in Supplementary Information (Figure S1-S23, Supporting Information).Printed tools allowed the robot to use various standard chemical consumables (tubes, tips) and equipment (automatic pipette, vacuum filtration system, centrifuge) that are usual for each chemical laboratory.Additionally, the laboratory station's physical drawing is represented in Figure S24 (Supporting Information).
The required GO and PEI solution amounts are mixed in a 50 mL laboratory tube.The process goes as follows: the right arm takes a tube from the tube storing bank and transfers it to a tube opening station, while the left arm removes a 100 mL bottle cover and brings the bottle with PEI solution to a dosage station.Afterward, while the left arm is holding the tube in the station the right arm removes the stationed tube's cap and places it in the designated position.The tube is transported to its according position in a dosage station by the left arm while the right arm retrieves the pipette dispenser and equips a 5 mL tip.The required amount of solution dosages is calculated according to the part volume.After dispensing the last part, the right arm moves the dispenser to a nozzle removal position.The left arm presses the nozzle ejection button on the dispenser and then manually grabs the ejected nozzle to avoid uncontrolled ejection.The right arm brings the dispenser back to its base while the left one transports the used nozzle into a waste disposal container.
The left arm brings the PEI solution bottle back into the bottle bank, sets the cover on top of the bottle, and then brings the GO solution to the dispensing base, repeating the steps mentioned above.After the dispensing process is done, the tube is moved back to a tube-closing opening station to tightly screw the cap back on the tube utilizing both robotic arms.Then, the left arm lifts and shakes the solution for 10 min to mix both solutions properly.After each centrifugation step, this process is modified due to the residue-washing requirement.Instead of GO and PEI solution, 0.1 NaCl solution is poured into the retrieved from the centrifuge tube, but the basic actions remain the same.During this process, two tubes containing solution are prepared to balance the centrifuge for safety reasons.Artificial vision is used for operation with a centrifuge.
After the reactive mixing process, the tube must be centrifuged for 90 min at 4000 RPM speed.The left arm presses the "open" button on the centrifuge control panel and lifts the centrifuge's lid.A computer vision system based on a COGNEX camera is utilized to find the centrifuge's rotor position.The pattern the camera tries to find is presented as a contrast sign on top of the rotor.After successful position detection, the left arm brings the rotor into a set position and places two opposite tubes into the corresponding positions.Then, it closes the lid and presses the centrifugation "start" button.After 90 min, the robot opens the centrifuge again and retrieves one of the tubes for further washing, described as a part of the previous step.
The right arm picks a filter from the blank filter box utilizing a vacuum suction cup and sets it on top of the vacuum filtration system.When the centrifugation and washing processes are completed, the filtration system assembly is performed by placing and holding a funnel on top of the flask.The left arm turns on the filtration system's vacuum pump and pours the resulting mixture into the funnel.The filtration takes 30 min, after which the robot turns off the pump and returns the funnel to its respective stand.
When the filtration vacuum system is disassembled, the right arm takes a Petri dish from the appropriate stand.It brings it closer to the vacuum filter, where the resulting membrane resides after the previous step.The left arm picks tweezers from the stand, carefully picks the membrane from the top of the filtration flask, and places it into the prepared petri dish.The right arm waits until the exicator cabinet is opened by the left arm, after which it puts the Petri dish into the cabinet for 24 h drying process.
After 24 h drying period, the testing stand needs to be assembled to conduct analysis.To achieve it, the right arm takes two analysis bottles and sets them onto an analysis stand on the stationary and mobile platform.The left arm opens the dry cabinet to retrieve the analyzed membrane.Afterward, the right arm picks a Petri dish with the required dried membrane and sets it on top of the vacuum filtration flask.Then, the right arm picks a membrane and puts it into a specific position on the analysis stand close to the stationary bottle utilizing the vacuum suction cup.When both bottles are placed in their positions, and the membrane is set between them, the robot sends a digital signal to the Arduino Uno board that controls the NEMA 17-step motor to bring both bottles together.Utilizing similar actions as in a mixture of the solution process, the robot fills one bottle with the tested solution and the other with sucrose.After 24 h of using a pipette dispenser and mobile pipette stand near the analysis cell, the robot retrieves the solution's aliquot from each reservoir into the laboratory tube.The robot closes the tubes and sets them aside to be collected by the researcher later.Then, the robot disassembles the analysis stand by resetting the control digital output and removes the bottles after both reservoirs reach their initial position.The membrane prepared by Cobot demonstrated the same morphology without any wrinkles or other features on the surface (Figure 2a,b).The pH of the polyelectrolytes and analytical solutions was variated during the experiments.28] Reverse osmosis (RO) membrane testing setup was developed to make a continuous process of permeability measurements.The system allowed us to test three membranes simultaneously and monitor the feed and permeate's pH and conductivity over time.Four salts (LiCl, NaCl, KCl, CsCl) were chosen for the measurements.Manually, in the RO setup, pH of the feed was adjusted from 2 to 10 (Figure 2).
Each step led to polyelectrolyte conformation changing, which impacted the ionic channels. [29,30]Figure 2d-k demonstrates a change in dependency of pH and conductivity after membrane over time. [31]As reported before, [21] the GO-PEI membrane closed for Cs þ , Li þ , and Na þ , corresponding with low and invariable for the first time (5-18 h).Surprisingly, conductivity in the case of CsCl increased after 18 h.We hypothesize that the membrane swelled critical concentration of protons (H þ ) and started to pass them (changing the pH of the permeate) simultaneously with Cs þ cations.For Li þ and Na þ cations, such an effect was not detected.We supposed that the presence of GO and PEI ratio creates an ionic channel only for K þ and Cs þ cations (in some cases).However, further research could help adjust the membrane for other cations transfer by the compound or changing environment.
A ML technique was applied to analyze data on permeability for a mix of CsCl and KCl (Figure 3).Here, we used the same analysis method as in the previous part.These salts were chosen to prove the effect of joined transfer through the membrane during the time.Unexpected permeate conductivity decreased over time (Figure 3e).It could be explained by the cations' competition in membrane layers, which can stack cations inside.
Figure 3b represents an outcome of recurrent neural network (RNN) predictions based on partial initial data, and the bottom graph shows the relation between the whole dataset and RNNpredicted pH values.As the bottom graph proves, the trained RNN's prediction accuracy is 88%.The database was divided into training, validation, and test samples in the 60, 20, and 20 ratios.Data from a specific sample from the selected time series were fed to the model input.The model's architecture represented one long short-term memory (LSTM) module and a linear activation function (Figure 3a).L1Loss chose the loss function. [32]The number of hidden layers is 5, and the number of neurons on  [25] Copyright 2021, Nature.c) Schematic illustration of permeability process dependence of media pH.d,e) Aqueous solutions 0.1 M KCl, f,g) 0.1 M LiCl, h,i) 0.1 M CsCl, and j,k) 0.1 M NaCl were tested with pH changing from base to acid values by manually addition of 0.1 M HCl.Conductivity and pH were measured for feed (black curves) and permeate (red curves) volumes.
each layer is 128.The most suitable parameters were determined by optimization using the Optuna framework.
The minimum loss function was found using the Adam optimizer, and the base learning rate was set to 10 À2 .Scheduler -Reduce on Plateau, which reduces the learning rate if there is no change in the selected metric for a certain number of epochs.The methodology was chosen losses on the validation sample, the minimum value of the learning rate is 10-8, and the number of periods without changing is 10.The Early Stopping method was also used to stop the learning process early if the local minimum of the loss function was reached.The number of epochs for an early stop is 20.The resulting model was evaluated using the metrics mean absolute error, mean squared error, and R2 (Figure 3c) on a test sample that did not participate in the training.Figure 3d,e visualizes the real-time series for the pH dependence on conductivity and the predicted time series.The data obtained indicate that the model effectively predicts the values of the time series.

Conclusion
Present research highlights the importance of the chemical processes robotization and deep learning algorithms in developing interpretable ML models of membrane structure.In terms of membrane synthesis automation, the full cycle to synthesize two membranes took 4 full hours to finish, which is 1 h longer than the manual membrane synthesis.However, the platform can perform manipulations with the laboratory equipment while there are enough solutions to experiment, in our case, the time of autonomous work was defined as 24 h.Therefore, the period of autonomous work depends solely on material provision, which can be automated to bring us one step closer to fully autonomous laboratory realization.This results in fully understanding the fundamentals of permeability of polyelectrolyte-based membranes and developing strategies for perspective filtration systems.The RNN predicted a GO-PEI membrane pH by permeate conductivity.The discussed model can be used for real-time pH value prediction and synthesis parameters adjustments to obtain the membrane with the required properties.Separate training is required for each variation of the synthesis process, however, with the increasing amount of data obtained by the automated platform, it is possible to create a universal model capable of indicator correction and possible solutions ratios and combinations prediction to synthesize a membrane with specific properties.

Experimental Section
Robot Description: The experimental setup utilizes a two-handed ABB YuMi collaborative robot mounted on a stationary platform.Each robot arm has a rated payload of 0.5 kg, a reach of 800 mm, and a working range of 559 mm.The platform is comprised of an aluminum profile on which the robot is mounted on through a 10 mm chipboard tabletop.The robot has a mass of 38 kg and combined with a stationary platform; the experimental setup mass reaches around 100 kg, excluding workstations placed on the tabletop.Gripper fingers were specifically designed to handle 50 mL laboratory tubes, 100 mL bottles, and Petri dishes and enable the robot in the vacuum filter assembly process.Vacuum suction cups were installed on the right arm's gripper for membrane and Petri dish handling.The vacuum compressor with the receiver was placed under the stationary platform for the pressurized air supply of the built-in robot vacuum system.
Experimental Setup: The automated membrane synthesis and analysis process is described in five major steps: liquid dispensing, centrifuge handling, vacuum filtration, membrane desiccation, and analysis.The first and second steps are repeatable throughout the process due to membrane synthesis technology.The liquid dispensing step involves several workstations, such as a tube holder, pipette dispenser, replaceable pipette holder, tube opening station, and dispensing base.Most of the banks and supplementary stations were made using fused deposition modeling (FDM) 3D printing technologies.A specific modification has been made to make use of the pipette dispenser due to the collaborative nature of the robot.It was encased into the modified handle, which eases the process of the dispenser handling by the robot.To reduce the amount of occupied space on the tabletop, the tube holder was designed to provide the required number of tubes to supply 24 h work cycle.Several workstations, such as the Eppendorf centrifuge 5804, vacuum filter pump, and dry cabinet, placed on top of the robot were used to ensure a stable workflow of membrane synthesis.The supplementary system composed of a NEMA17 step motor and Arduino UNO board with a motor shield was included to provide the membrane analysis data acquisition process, utilizing the single arm of the robot.
Membrane Preparation: Membranes were prepared by vacuum filtration of GO and PEI on a polyethersulfone filter.Water flux was investigated by using an H-type sealed two-compartment electrolytic cell of 50 mL for 24 h.2.5 M sucrose solution was added in one cell, and water solution with a particular pH was added in the second half. [25]obot Control: The programming code for each hand was written to ensure the stable completion of the work cycle for each hand in the RobotStudio development environment.For each procedure, a series of waypoints comprising XYZ coordinates relative to the robot's center of coordinates and the orientation of each link in quaternions were set.Global Boolean variables were used to prevent work cycle interruptions due to hand collision.For repetitive tasks with a variable waypoint, such as membrane handling, pipette nozzle equipping, Petri dish handling, etc. some points were derived by the transition of the workstation base point.To ensure stable completion of the centrifuge handling task, a computer vision system utilizing a Cognex camera was mounted on a dry cabinet for tube position finding inside the centrifuge.
The RO Membrane Testing Setup: RO setup was developed to automatically measure permeability through membranes by conductivity and pH electrode placed into the permeate solutions (Figure S1, Supporting Information).User-prepared membranes tested against a typical feeding solution.
Prediction: A RNN was introduced into the robotic complex to predict membrane properties based on desirable filtration outcomes.We have used Python version 3.9 with the PyTorch framework and pytorch_lightning extension to build the model with LSTM architecture.NumPy library for linear algebra, Pandas library for working with tabular data, Scikit-Learn library for data processing, and Matplotlib library for visualization were also used.This approach proved to be the most reliable in terms of accuracy for handling time-stamped data.The models were trained on a personal computer with an AMD THREADRYPER 3960X 24-core 3.9 GHz CPU, 64 GB RAM and two NVIDIA RTX 3090 GPUs.Training and optimization took about 2 h.A personal computer with an Intel Core i5-10600 CPU @ 3.30 GHz processor, 32 GB of RAM, and a GeForce RTX 2060 video card was used to train the RNN model.RNN model was taught on a case of 0.1 KCL and 0.1 CsCL separate and combined solutions filtration by GO-PEI membranes.The permeate conductivity and pH datasets were divided into training and validation datasets using random dispersion with an 80%-20% train-eval relation.

Figure 2 .
Figure 2. a) Photo of membrane synthesis by robot.b) Scanning electron microscope images of GO-PEI membrane.Reproduced with permission.[25]Copyright 2021, Nature.c) Schematic illustration of permeability process dependence of media pH.d,e) Aqueous solutions 0.1 M KCl, f,g) 0.1 M LiCl, h,i) 0.1 M CsCl, and j,k) 0.1 M NaCl were tested with pH changing from base to acid values by manually addition of 0.1 M HCl.Conductivity and pH were measured for feed (black curves) and permeate (red curves) volumes.

Figure 3 .
Figure 3. a) RNN model structure with LSTM cell scheme.b) RNN prediction pH outcome and real data comparison.c) pH prediction based on partial initial data.d) pH and conductivity and e) measurements for a mix of 0.1 M KCl and 0.1 M CsCl.